Xiao Liu PhD Thesis.pdf - Faculty of Information and Communication ...
Xiao Liu PhD Thesis.pdf - Faculty of Information and Communication ...
Xiao Liu PhD Thesis.pdf - Faculty of Information and Communication ...
Create successful ePaper yourself
Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.
Thus, the sufficiency property holds.<br />
Here, we also adopt the method <strong>of</strong> DOMTR (Dynamic Obtaining <strong>of</strong> Minimum<br />
Time Redundancy) [21]. Based on some initial values which are set up during the<br />
runtime instantiation stage, DOMTR can compute the minimum probability time<br />
redundancy on the fly along scientific workflow execution with minimal<br />
computation. Based on DOMTR, the computation <strong>of</strong> our checkpoint selection<br />
strategy is basically one or two subtractions or comparisons at each activity covered<br />
by one or more upper bound constraints. Therefore, as proved in [21], the<br />
computation cost is basically negligible.<br />
To conclude, our checkpoint selection strategy strictly ensures that a checkpoint<br />
is selected only when current temporal consistency is below the minimum<br />
acceptable threshold. Meanwhile, since our temporal verification strategy is aware<br />
<strong>of</strong> the effective probability consistency range <strong>and</strong> can determine the fine-grained<br />
levels <strong>of</strong> temporal violations, it also ensures that statistically the time deficit can be<br />
compensated by light-weight temporal violation h<strong>and</strong>ling.<br />
6.3 Evaluation<br />
6.3.1 Experimental Settings<br />
The SwinDeW-C simulation environment is introduced in Section 3.5 <strong>and</strong> hence<br />
omitted here. The experiment settings are as follows. The process definitions are<br />
generated based on the pulsar searching example presented in Section 1.2 <strong>and</strong> the<br />
attribute settings are based on historic data for the running <strong>of</strong> scientific workflow<br />
instances [56]. Meanwhile, the experiment settings are extended deliberately for the<br />
purpose <strong>of</strong> exploring a much larger parameter space to evaluate our strategy in<br />
general s<strong>of</strong>tware applications <strong>of</strong> similar nature in distributed high performance<br />
computing environments.<br />
To evaluate the average performance, 10 independent experiments with different<br />
scientific workflow sizes (ranging from 2,000 to 50,000 workflow activities with<br />
their mean activity durations between 30 to 3,000 basic time units 4 ) are executed<br />
4 Here, a basic time unit is used for general evaluation purpose. Specifically, for our pulsar searching<br />
example, a basic time unit here is more or less equal to a second.<br />
98